Joe Biden will win the US election according to a technique used in finance to predict factor returns and the correlation of stock and bond returns.
The past as a prologue: how to forecast presidential elections, a MIT working paper, co-authored by MIT’s Mark Kritzman and Dave Turkington and Megan Czsasonis from State Street, uses a model that correctly predicted the outcomes of the past five presidential elections. It is now predicting a strong Democratic victory for 2020, with a Democratic loss within a confidence band of one standard deviation. Interestingly, the technique correctly predicted the 2016 election, which the polls failed to do.
The authors apply a novel forecasting technique called Partial Sample Regression which measures the statistical relevance of past elections, and then employs an “obscure mathematical equivalence” – that the prediction from a linear regression equation equals a relevance-weighted average of the values for the dependent variable. It uses this to forecast election outcomes from a subsample of prior relevant elections.
The technique predicts the elections in a mathematically formal way and uses no poll data.
“The essence of our methodology is to measure the relevance of historical elections in a statistically rigorous way. We then rely on an obscure mathematical equivalence to form predictions from the more relevant elections,” the authors say.
“When political scientists or pundits forecast presidential elections, they often analyse past elections for clues about upcoming elections. But they don’t treat all past elections alike. They judge some to be more relevant than others. This behaviour is true in general when we try to predict an outcome based on prior experiences. We look for those events that bear some resemblance to current conditions. We apply this concept to predict the outcomes of presidential elections, but we do so in a mathematically formal way.”
The authors then add to this the less obvious component of “relevance”.
“We consider the unusualness of the past experiences. The intuition is that unusual occurrences are more informative than common occurrences, which simply might be a manifestation of noise in the data. Once we identify a subsample of relevant historical elections, we invoke an obscure mathematical equivalence. The prediction from a linear regression equation equals a weighted average of the past values of the dependent variable in which the weights are the relevance of the values for the independent variables. We apply this equivalence to our relevant subsample of political, geopolitical, and economic data to form our predictions.”
The methodology used to predict election results is similar to that used by Kritzman and Turkington and their co-authors Ding Li and Grace Qiu from GIC in the paper, Portfolio Choice with Path Dependent Preferences, which is forthcoming in the Financial Analysts Journal. That study, which revolutionises scenario analysis by reorienting it towards a path rather than a single period outcome, finds that a U-shaped recovery is the most likely economic outcome in the US for the next two years, but stagflation has a higher than anticipated chance of occurring.